Guardrailz
HomeHubBlogsPricingDocs

Introduction

  • Overview
  • Architecture
  • Core Concepts

Getting Started

  • Installation
  • Quickstart
  • Playground

API

  • Validate API
  • Profiles API
  • Analytics API
  • Error Handling

SDK

  • Overview
  • Client
  • Guardrails
  • Examples

Guardrails

  • Overview
  • Input Guardrails
  • Output Guardrails
  • Tool Guardrails
  • Custom Guardrails

Profiles

  • Overview
  • Built-in Profiles
  • Custom Profiles
  • Profile Compilation

Analytics

  • Overview
  • Events
  • Queries
  • Dashboards

Deployment

  • Environment
  • Security
  • Scaling
  1. Docs
  2. /introduction
  3. /concepts

Core Concepts

This section introduces the key concepts you need to understand when working with Guardrails.

These concepts form the mental model of the system.

Guardrail

A guardrail is a unit of logic that evaluates content or context and produces a structured result.

A guardrail:

  • Has a name and stage
  • Receives input and context
  • Returns a decision (allow, warn, block, modify)
  • Can attach metadata

Examples:

  • Input size limits
  • Prompt injection detection
  • PII redaction
  • Tool access control

Guardrail stages

Guardrails run in specific stages:

  • Input – before the model is invoked
  • Output – after the model produces output
  • Tool – when tools or actions are executed
  • General – cross-cutting checks

Stages allow precise control over when enforcement happens.

Profile

A profile is a collection of guardrails and configurations.

Profiles allow you to:

  • Apply consistent policies
  • Switch behavior by use case
  • Manage complexity at scale

Profiles can be:

  • Built-in
  • Organization-specific
  • User-specific
  • Environment-specific

Guardrail result

Each guardrail returns a structured result containing:

  • Pass/fail status
  • Action taken
  • Severity level
  • Optional message
  • Optional metadata

Results are aggregated into a single execution response.

Execution context

The execution context provides additional information to guardrails, such as:

  • User identifiers
  • API key identifiers
  • Profile identifiers
  • Validation type (input/output)
  • Tool invocation data

This enables context-aware enforcement.

Analytics event

An analytics event captures what happened during execution.

Events include:

  • Guardrail execution events
  • Profile usage events
  • Rate-limit events

Events power:

  • Dashboards
  • Alerts
  • Compliance reporting

SDK client

The SDK client is the developer-facing interface for interacting with Guardrails.

It abstracts:

  • API calls
  • Authentication
  • Error handling
  • Retries and timeouts

Putting it together

In practice:

  1. A request is received
  2. A profile is resolved
  3. Guardrails are executed
  4. Results are enforced
  5. Analytics are recorded

This flow repeats consistently across all use cases.

Next steps

  • Try Guardrails quickly → Quickstart
  • Learn guardrails in depth → Guardrails Overview
  • Explore profiles → Profiles Overview